How to Convince Someone That You Can Be Trusted?

The Role of ‘Hostages’
Chris Snijders and Vincent Buskens

Department of Sociology
Utrecht University
March 22, 2001
1 Introduction
Sometimes, cooperation can be mutually profitable and can be foreseen to run smoothly.
However, sometimes cooperation will not materialize because the actors involved an-
ticipate trouble along the route to mutual profit. As our main paradigm representing
a potentially mutually profitable enterprise between two actors, we consider the Trust
Game (Dasgupta 1988; Kreps 1990; Snijders 1996). Figure 1 shows the Trust Game in
its general format. A lack of trust renders players 1 and 2 a payoff (P
1
,P
2
), whereas trust
that is honored will result in the Pareto-improvement (R
1
,R
2
). Unfortunately, player 2 is
tempted to abuse the trust given to him because that would give player 2 an even larger
payoff of T
2
. Foreseeing the abuse of player 2, which would leave player 1 with only S
1
,
player 1 might well decide not to trust player 2, which in turn results in both players
getting the Pareto-inferior (P
1
,P
2
) outcome.
There are many conditions under which cooperation problems like in the Trust Game
are less severe or even completely absent. For instance, when the interaction is guarded
by norms of altruism or reciprocity (Gouldner 1960), when interactions are embedded
over time, or when the existence of social networks ensures cooperative behavior through
the effects of reputation (Raub and Weesie 1990; Buskens 1999). In this contribution, we
restrict our attention to posting hostages in the sense of Schelling (1960) as a solution to
the cooperation problem. Although the word ‘hostage’ may cause confusion since it can
be associated with ‘hostage taking’ (which is not what we are dealing with), we maintain
this terminology as as a tribute to Schelling’s original work (cf. Schelling 1960, since
he was (one of) the first to recognize the general mechanism of reshaping the incentive
structure to attain a mutually profitable outcome. Also in more recent work he repeatedly
emphasized the usefulness of hostage posting, especially in cases in which trust is essential
to the interaction. In his (1984) words:

Stimulating comments of and discussions with Werner Raub and Jeroen Weesie are gratefully ac-
knowledged. Financial support was provided by the Netherlands Organization for Scientific Research
(NWO) under grant PGS 50-370 and the Royal Netherlands Academy of Arts and Sciences (Snijders) .
Address: Department of Sociology, Utrecht University, Heidelberglaan 1, 3584 CS Utrecht, the Nether-
lands. Email: c.snijders@fss.uu.nl.
1
Player 1
Player 2
?
-
-
-
P
1
P
2
S
1
T
2
R
1
R
2
Player 1Player 2
trust
honor
abuse
no trust
Figure 1: The Trust Game: S
1
< P
1
< R
1
, P
2
< R
2
< T
2
.
“(...) hostages (...) are techniques for structuring incentives to assure compli-
ance. And when trust or confidence is essential to an enterprise, the hostages
may be given voluntarily. (...) The procedure may be demanded or volun-
teered; when volunteered, it is not altogether different from the voluntary
jeopardy that people incur when they post bail or pawn a camera.”
What prevents the mutually cooperative outcome to emerge, is not so much player 2’s
incentive to abuse trust per se, but it is player 1 anticipating the abuse of trust by player 2.
This is why it is also in player 2’s interest to convince player 1 that he will not abuse trust
if it is given. One way to try to convince player 1, would be for player 2 to (voluntarily)
decrease (or remove) the incentive to abuse trust. This mechanism—decreasing one’s own
incentives—is how we operationalize Schelling’s “hostage posting”. Using experiments on
Trust Games, we consider the conditions under which hostage posting in the Trust Game
solves the trust problem.
2 Theory
To see how a hostage can change the incentive structure, consider the game in Figure 2.
Player 2 has to decide whether or not to restrict his own future incentives by posting
a hostage. Posting a hostage creates a change in the incentive structure by imposing a
sanction on abusing trust. If the hostage is not posted, player 1 and player 2 play a
Trust Game Γ
TG
with monetary payoffs (S
1
,P
1
,P
2
,R
1
,R
2
,T
2
). If player 2 posts a hostage
of monetary value H, then player 1 and player 2 play a Trust Game Γ

TG
with monetary
payoffs (S

1
,P
1
,P
2
,R
1
,R
2
,T

2
), where S

1
= S
1
+ H
1
and T

= T
2
−H
2
. The hostage has an
impact on the monetary payoffs of the underlying Γ
TG
only in the case that player 2 has
posted the hostage. Suppose that player 1 decides to trust player 2, and subsequently
player 2 abuses player 1’s trust. In that case, player 2 gets T
2
− H
2
(instead of T
2
), and
player 1 gets S
1
+ H
1
(instead of S
1
). This way, the hostage not only puts a penalty on
abusing trust for player 2, but it also gives compensation for abused trust to player 1 (if
2
Player 1
Player 2
?
-
-
-
P
1
P
2
S
1
+H
1
T
2
−H
2
R
1
R
2
P
1
P
2
S
1
T
2
R
1
R
2
Player 1 Player 2
DOWN
DOWN
RIGHT
RIGHT
Player 1
Player 2
?
-
-
-
DOWN
DOWN
RIGHT
RIGHT
Player 2
Post
hostage
Do not
post hostage
Figure 2: Hostage Trust Game Γ
HTG
.
H
1
> 0).
1
For obvious reasons, the Trust Game with a hostage option is called a Hostage
Trust Game. Whereas a Trust Game is defined by the values (S
1
, P
1
, P
2
, R
1
, R
2
, T
2
), a
Hostage Trust Game Γ
HTG
is defined by (S
1
, P
1
, P
2
, R
1
, R
2
, T
2
; H
1
, H
2
).
A hostage in a Trust Game solves the trust problem if four conditions are fulfilled. (1)
Player 1 would not be willing to trust player 2 in a certain Trust Game Γ
TG
; (2) player 1
would also not trust player 2 if player 2 had not posted the hostage in the Hostage Trust
Game Γ
HTG
; (3) player 1 trusts player 2 after player 2 has posted the hostage; and (4)
player 2 honors trust after posting the hostage and after player 1 has trusted player 2.
In short, a hostage solves a problem of trust, if the hostage convinces player 1 to trust
player 2 after the hostage has been posted, and player 1’s trust is justified: player 2 honors
trust.
Two kinds of hostages are—from an analytical point of view—less interesting, even
though they are likely to solve a problem of trust. Hostages that fully compensate player 1
in case his trust gets abused and hostages that completely bind a player 2 to honor trust.
In Figure 2, consider the game that arises after player 2 has posted the hostage. The
subsequent game after a hostage has been posted that fully compensates player 1 in case
of abused trust (H
1
> P
1
− S
1
) is no longer a Trust Game in monetary terms. The
1
All hostages are assumed to fulfill H
2
≥ 0. This specification disregards the expropriation hazard
(Becker 1991: 12–13,Williamson 1985: 177): a posted hostage can only be lost if player 2 has indeed
abused trust.
3
problem of trust is definitely “solved” in that case, as long as utility is assumed to be
increasing in own monetary gains. In this case, there is no longer any need to fear being
cheated as player 1. We call this a bribing hostage. After a hostage has been posted that
removes the incentive to abuse trust for player 2 (H
2
> T
2
− R
2
), the subsequent game
is also no longer a Trust Game in monetary terms. Any sensible player 2 would honor
trust after such a hostage was posted. Unless player 1 thinks that player 2 is not that
sensible (which sometimes seems to be the case, cf. Rosenthal 1981; Beard and Bell Jr
1994; Keren and Raub 1993), player 1 will trust a player 2 who posts such a hostage. As
long as player 2’s utility function is increasing in his own monetary gains (and player 1
knows it), the problem of trust has also been solved. We call this a completely binding
hostage. We do not consider hostages that are either bribing or completely binding.
2.1 Convincing player 1
2.1.1 Results from Trust Games - the behavior of player 1
If we consider only those hostages that do not fully compensate player 1 and do not
completely bind player 2 (i.e, hostages with H
2
< T
2
− R
2
and H
1
< P
1
− S
1
), we
can ‘derive’ hypotheses based on results from experiments with Trust Games. Assuming
that utility is a function of ego’s monetary payoff and alter’s monetary payoff, one can
deduce how behavior in Trust Games should depend on (functions of) the separate payoffs
(see Snijders 1996: Chapter 4). In Trust Games, one can show both theoretically and
empirically that risk =
P
1
−S
1
R
1
−S
1
and temptation =
T
2
−R
2
T
2
−S
1
have a negative effect on the
probability that player 1 trusts player 2. Moreover, empirically one can show that the
effect of risk on this probability is larger than the effect of temptation =
T
2
−R
2
T
2
−S
1
. Loosely
translated, one can say that to player 1 it matters more what he can end up with himself
(S
1
, P
1
, or R
1
) than what player 1 thinks player 2 is likely to do. If these are results
that carry through to Hostage Trust Games, we expect negative effects of the degree of
compensation
H
1
P
1
−S
1
and the degree of binding
H
2
T
2
−R
2
on the probability that player 1
trusts player 2. Furthermore, the effect of a change in payoffs for player 1 (compensation)
should have a larger impact on the probability that player 1 will trust player 2 than a
change in the payoffs for player 2 (binding).
Hypothesis 1 Compensation increases trust. The probability that play-
er 1 trusts player 2 increases with
H
1
P
1
−S
1
.
Hypothesis 2 Binding increases trust. The probability that player 1
trusts player 2 increases with
H
2
T
2
−R
2
.
Hypothesis 3 Compensation is more important than binding. Con-
sider the Hostage Trust Game Γ
HTG
. With respect to the probability that
player 1 trusts player 2, the degree of compensation a hostage provides (to
player 1), is more important than the degree of binding a hostage provides (to
player 2).
Suppose player 2 posts the hostage. There are two ways in which player 2’s decision
may affect player 1’s subsequent behavior. Player 1 can decide to trust player 2 because
the mere fact that player 2 posted the hostage signals that player 2 is the kind of player
4
that is going to honor trust in the Trust Game Γ

TG
. Player 1 then considers the hostage
to be type revealing. Or, player 1 can decide to trust player 2 because he considers it
likely enough that player 2 will honor trust in the Trust Game Γ

TG
even though player 1
may not have been willing to trust player 2 in the Trust Game Γ
TG
. Player 1 considers
the hostage to be binding. We argue that in the case of the Hostage Trust Game, under
reasonable conditions, type revealing hostages do not exist.
2.1.2 Type revealing hostages
There is little opportunity for hostages to be type revealing in the Hostage Trust Game
because posting a hostage is free of charge. Intuitively, it is hard to imagine how a
signal that is free of charge can reveal a characteristic about a player. After all, players
who do not possess that particular characteristic can pretend to possess it for free.
2
For
instance, a door-to-door salesman wants to appear reliable to a potential customer, and
could therefore decide to wear a suit. But, since all potential customers know that any
door-to-door salesman wants to appear reliable and will therefore wear a suit, the suit
is no longer a signal that distinguishes the reliable salesmen from the unreliable ones.
Something similar is going on here. Given reasonable interpretations of the interaction,
there are no real game-theoretic arguments for a hostage to fulfill the role of a credible
signal of benign intent.
3
To illustrate that claim, we apply a model that we call the guilt
model to the Hostage Trust Games. In the guilt model, a parameter γ
2
is introduced that
represents a property of player 2: it indicates the extent to which player 2 feels guilty as
a result of the larger payoff he obtains after abusing trust. We show that player 1 cannot
deduce the value of γ
2
from the fact that player 2 posted the hostage.
Considering only those hostages that are not bribing and not completely binding
(H
1
< P
1
− S
1
and H
2
< T
2
− R
2
) implies that only those hostages are considered
that—if posted—lead to a game that is still a Trust Game in monetary terms. We apply
a model of man that assumes “guilt” in case trust is abused. In the guilt model, utility is
defined as U
i
(M
i
, M
j
) = M
i
−γ
i
max(0, M
i
−M
j
). This implies in the Hostage Trust Game
that the utility of player 2 is affected by his parameter γ
2
. The trust parameter indicates
the extent to which player 2 feels guilty because he deceives player 1. Straightforward
game-theoretic analysis (see Snijders 1996; Snijders and Keren 1999) shows that player 2
will honor trust in the part of the Hostage Trust Game without a hostage if his trust
parameter γ
2
>
T
2
−R
2
T
2
−S
1
and, correspondingly, in the part of the game with a hostage if
γ
2
>
T
2
−R
2
−H
2
T
2
−S
1
−H
1
−H
2
. Figure 3 shows the Hostage Trust Game in “guilt model format”.
If player 2 does not post a hostage, the subsequent game is again the original Trust
Game Γ
TG
, only now in its guilt model format. The upper part of the figure shows that
after player 2 does post a hostage, the subsequent game is a Trust Game Γ

TG
, also in
its guilt model format. That is, depending on whether or not player 2 posts the hostage,
player 2 can enter one of two Trust Games. Player 2’s task is to decide which of these two
2
Obviously, this argument also goes for signals that are sufficiently cheap. On the other hand, “free
signals” can of course affect the incentive structure in games of coordination. If two persons want to meet
somewhere in Paris and decide to meet under the Eiffel Tower, there is no need for them to actually go
there, but both have no incentive to deviate from the agreed place. Benchmark articles about the role of
“cheap talk” in communication are Crawford and Sobel (1982, Farrell (1987).
3
Note the “game-theoretic” in this sentence. There is nothing against assuming, for instance, that the
sole posting of a hostage by player 2 triggers trustfulness in player 1. This may be a valid argument (in
fact, the data will support it) but it is not a game-theoretic argument.
5
l
l
2
1
Player 1 Player 2
R
1
S
1
+H
1
P
1
R
2
T
2
−H
2
−γ
2
|T
2
−S
1
−H
1
−H
2
|
P
2
R
1
S
1
P
1
R
2
T
2
−γ
2
(T
2
−S
1
)
P
2
l
l
2
1
l
2
Post
Hostage
Do not post
Hostage
Trust
Trust
¬ Trust
¬ Trust
Honor
Honor
¬ Honor
¬ Honor
Figure 3: Hostage Trust Game Γ
HTG
in guilt model format.
Trust Games he would prefer to be in, while taking into account that whether player 1
will choose to trust player 2 may well depend on player 2’s hostage posting decision.
The appropriate game-theoretic format is a game of incomplete information. Actually,
for analytical purposes, Figure 3 is not the correct way to proceed. We should have
depicted a game tree in which “Nature” moves first and chooses a value γ
2
for player 2.
Whereas the distribution out of which the γ
2
was chosen is assumed to be common
knowledge, only player 2 knows its value. In principle, the posting of a hostage could
reveal something to player 1 about the value of γ
2
. For instance, an extreme case would
be the one in which player 2 follows one of two possible strategies: either player 2 posts and
subsequently honors trust, or player 2 does not post a hostage and will not honor trust. In
such a case (a “separating equilibrium”), if player 2 posts a hostage, this unambiguously
displays the fact to player 1 that he can trust player 2. However, something of that nature
can never occur in the Hostage Trust Game in its guilt model format.
To see why player 1 can never infer something about γ
2
after the hostage has (or
has not) been posted, suppose that player 1 trusts player 2 if and only if player 2 has
posted a hostage. Player 2’s decision then reduces to either not posting the hostage and
getting P, or posting the hostage and getting at least R. Hence, irrespective of the value
of γ
2
, player 2 will decide to post the hostage, and player 1 cannot infer anything about
the value of γ
2
on the basis of player 2 posting a hostage. Allowing for the use of more
sophisticated strategies, for instance, strategies in which player 1 decides to trust player 2
with probability q
h
after player 2 posts a hostage, and trusts player 2 with probability q
n
(q
h
> q
n
) otherwise, does not lead to different results (see Snijders 1996: Appendix B).
Likewise, it could be argued that since posting a hostage does not give player 1 any
information about the value of γ
2
, posting no hostage also does not provide player 1 with
information about γ
2
.
4
This immediately implies the next hypothesis.
4
Not posting the hostage is not part of the game-theoretic equilibrium. It is therefore not obvious
what player 1 should think given that player 2 declines to post the hostage. We assume that when
6
Hypothesis 4 Not posting a hostage has no negative side effects.
Consider a Trust Game Γ
TG
and a related Hostage Trust Game Γ
HTG
with a
hostage that is not bribing and not completely binding (H
1
< P
1
− S
1
and
H
2
< T
2
− R
2
). Player 1’s decision whether or not to trust player 2 in the
Trust Game Γ
TG
is the same as player 1’s decision whether or not to trust
player 2 in the Hostage Trust Game Γ
HTG
after player 2 has declined to post
the hostage. In other words, the mere fact that player 2 has not posted the
hostage in the Hostage Trust Game, should have no additional influence.
The previous hypothesis compares player 1’s decision in a Trust Game with player 1’s
decision in a (related) Hostage Trust Game. Based on the same argument as above,
another sensible comparison could be made. Again, consider Figure 3. Compare the
two Trust Games Γ
TG
and Γ

TG
that arise as subgames in that Hostage Trust Game,
dependent on player 2’s decision to post the hostage. Define p
2,1
as the subjectively
expected probability that player 2 will honor trust, as assessed by player 1. If nothing
can be deduced from the fact that player 2 posts the hostage, player 1 will trust player 2
after no hostage has been posted if (and only if) p
2,1
>
P
1
−S
1
R
1
−S
1
, and player 1 will trust
player 2 after a hostage has been posted if (and only if) p
2,1
>
P
1
−S
1
−H
1

2
(T
2
−S
1
−H
1
−H
2
)
R
1
−S
1
−H
1

2
(T
2
−S
1
−H
1
−H
2
)
.
In other words, whether player 1 trusts player 2 is solely dependent on these two indices.
No additional effect of the posting of a hostage can exist. Rephrased as a hypothesis:
Hypothesis 5 Posting a hostage has no positive side effects. Consider
the Hostage Trust Game Γ
HTG
with a hostage H that is not bribing and not
completely binding (H
1
< P
1
−S
1
and H
2
< T
2
−R
2
). Compare the two Trust
Games player 1 can end up in, after player 2 has decided whether or not to
post the hostage. Player 1’s decision whether or not to trust player 2 in these
two Trust Games, is solely based on the payoffs that are brought about by
player 2’s decision. In other words, the mere fact that player 2 has posted the
hostage, has no additional influence.
2.1.3 Hostages that decrease the incentive to abuse
Though (theoretically) a hostage does not affect the decision of player 1 in the sense that
it suggests that player 2 is a trustworthy person, the hostage has of course a direct effect
on the monetary payoffs. The previous section showed that whether player 2 posts a
hostage, does not reveal any information about his guilt parameter. Hence, the subgames
that emerge after player 2’s decision whether or not to post a hostage, are Trust Games
that can be analyzed as such. We will analyze the effect of the hostages using the guilt
model. Note once again that the subgame that emerges if player 2 does not post the
hostage in a Hostage Trust Game Γ
HTG
, is the Trust Game Γ
TG
. The subgame that
emerges if player 2 does post the hostage, is the Trust Game Γ

TG
.
In the guilt model of Figure 3, player 1 trusts player 2 if and only if p
2,1
>
P
1
−S
1
R
1
−S
1
after player 2 chooses not to post the hostage. If player 2 did choose to post the hostage,
player 1 trusts player 2 if and only if p
2,1
>
P
1
−S
1
−H
1

2
(T
2
−S
1
−H
1
−H
2
)
R
1
−S
1
−H
1

2
(T
2
−S
1
−H
1
−H
2
)
. This immediately
such an “out of equilibrium move” occurs, player 1 does not alter his beliefs about the γ
2
of player 2.
According to the classification by Rasmusen (1994: 151), this assumption is called “passive conjectures”.
Cf. McLennan (1985) and Cho and Kreps (1987).
7
implies that player 1 will be “convinced” by the hostage if and only if both the previous
inequalities hold. In that case the hostage was essential in solving the problem of trust:
player 1 does not trust player 2 without the hostage, but does trust player 2 if the hostage
is posted.
Hypothesis 6 Conditions for hostages to be essential to establish
trust. Consider the Hostage Trust Game Γ
HTG
with a hostage that is not brib-
ing and not completely binding (H
1
< P
1
−S
1
and H
2
< T
2
−R
2
). A hostage
is essential when player 1 only trusts player 2 if a hostage is posted. This is
the case if (and only if) p
2,1
<
P
1
−S
1
R
1
−S
1
and p
2,1
>
P
1
−S
1
−H
1

2
(T
2
−S
1
−H
1
−H
2
)
R
1
−S
1
−H
1

2
(T
2
−S
1
−H
1
−H
2
)
.
2.2 The behavior of player 2
For a hostage to solve a trust problem, the hostage should not only convince player 1 to
trust player 2, but player 1’s trust should also be justified in the sense that player 2 is
now indeed more likely to honor given trust. The theory underlying player 2’s behavior is
much simpler than the theory for player 1. As soon as player 2 is trusted, player 2 plays a
standard Trust Game and will honor trust if (and only if) U(S
1
+H
1
, T
2
−H
2
) < U(R
1
, R
2
)
(if player 2 has not posted a hostage, both H
1
and H
2
equal zero).
2.2.1 Results from Trust Games - the behavior of player 2
In Trust Games, the probability that player 2 honors trust depends strongly on
T
2
−R
2
T
2
−S
1
(Snijders 1996: Chapter 4). Most of the variance in the behavior of player 2 is explained
by this variable (note that, perhaps surprisingly, the behavior of player 2 does not depend
on the value of P
2
). In the Hostage Trust Game case, player 2 will honor trust if (and only
if) γ
2
>
T
2
−R
2
−H
2
T
2
−S
1
−H
1
−H
2
if he did post the hostage, and if (and only if) γ
2
>
T
2
−R
2
T
2
−S
1
if he did
not post the hostage. Based on the game-theoretic argument deduced above that hostages
only have a direct effect on the monetary payoffs: Suppose that in a Hostage Trust Game,
player 2 has not posted the hostage, and player 1 subsequently trusts player 2. If it is
assumed that posting a hostage in itself does not create a moral imperative to honor trust,
player 2 is now in a similar position as in a Trust Game in which player 1 has decided to
trust him. If this is really the case, the behavior of player 2 must be likewise determined
by temptation.
Hypothesis 7 Hostage posting in itself is not a credible signal of
subsequent trustworthy behavior. Consider the Hostage Trust Game
Γ
HTG
with a hostage that is not bribing and not completely binding (H
1
<
P
1
− S
1
and H
2
< T
2
− R
2
). Player 2’s decision whether or not to honor
player 1’s trust in the Trust Game Γ
TG
is the same as player 2’s decision
whether or not to honor player 1’s trust in the Hostage Trust Game Γ
HTG
after player 2 has declined to post the hostage, and player 1 has decided to
trust player 2.
3 Data
We test our hypotheses using experimental data. An outline of the design follows below.
Additional details, including the complete instruction text, can be found in Snijders (1996:
8
Chapter 3).
Subjects
The study was conducted in two Dutch universities, the University of Amsterdam and
the University of Groningen. Subjects were recruited by campus advertisements and
were promised at least 10 Dutch guilders (and the possibility for a larger reward) for
participation. Nearly all participants were undergraduates from a variety of disciplines
with the exception of the study in Amsterdam in which 71 out of the 120 (59%) subjects
were students of Economics or Econometrics. The purpose of the latter sample was to
test whether these students respond differently since previous research (Frank, Gilovich,
and Regan 1993) suggests that they were more likely to behave in accordance with the
strict economic interpretation of game theory. Two subjects were removed from the data
set because it was clear from the questions they asked during the session that they did
not understand the essentials of the task. A few other subjects failed to make a choice
(e.g., one subject wrote “my choice depends on who player 2 will be”) or responded in an
unclear manner. Eventually, 8 out of a total of 234 subjects were eliminated from further
analysis.
Table 1: Overview of participants.
Groningen Amsterdam Total
Number of subjects 114 120 234
Rejected subjects 8 0 8
Number of subjects 106 120 226
Percentage of males 38 81 60
Average age 23 22 22
The subject population can be considered a more or less standard student population,
with the exception of the large percentage of students in economics or econometrics. A
second conspicuous aspect is the large proportion of males in the Amsterdam experiment.
This cannot be accounted for by the large proportion of students in econom(etr)ics, since
if only the non-econom(etr)ics students are considered, the large proportion of males
remains. Though we do not expect an effect of sex, it is something to keep in mind during
the analyses.
Procedure
Subjects were seated in a room with (about) fourteen other participants, always able to
see each other. The experimenter informed subjects (orally) that the experiment would
consist of a number of different tasks, some of which would concern choices regarding
monetary outcomes. They were further instructed to think carefully about their choices
because it was possible that (at the end of the session) some of them would be paid in
accordance with one of their choices. The procedure to determine the subjects who would
be paid and the specific game were explicitly outlined. First, one of the games was to
be randomly selected followed by a random selection of two subjects who would then be
paid according to their choices in that particular game. The procedure was explained by
means of a hypothetical example.
9
The purpose of applying such a procedure was to exclude “wealth-effects” (e.g., Kah-
neman and Tversky 1979; Davis and Holt 1993). Since each decision situation is selected
with equal probability, choices in each situation enter a subject’s overall expected utility
as additively separable components. Consequently, each component can be maximized
separately and potential earnings in one game should not affect decisions in other games.
The method is supposed to increase subjects’ motivation as there were considerable sums
of money at stake. The maximum amount a subject could earn was 120 Dutch guilders
(about 75 US$) which, according to student standards, is certainly a substantial amount.
Subjects were explicitly told that there were no “right” or “wrong” answers. They
were instructed to complete their tasks in the order in which they were presented in the
instruction text, and were given the possibility to ask the experimenter for assistance. In
each study, subjects performed three or four decision making tasks in which the Trust
Game task was the first one.
Tasks
Subjects were told that they were to complete several tasks: five Trust Games, a ques-
tionnaire, five Hostage Trust Games, and some other tasks not related to this experiment,
see Snijders 1996: 68).
The Trust Games were described (without mentioning concepts like trust or fairness)
by referring to a card with a display of a Trust Game in its extensive form. Subjects were
asked to determine whether they would trust player 2 in case they would be player 1,
and if they would honor trust of player 1 in case they would be player 2. These choices
constitute the (dummy) variables trust and honor trust.
For the Hostage Trust Games the instruction text pointed out that, again, subjects
had to make a decision as if they were player 1 and as if they were player 2 in certain
games. It was added that the game (and the rules) were the same as in the basic Trust
Game task, with one exception. Now, player 2 must make the first decision and must
decide whether or not to post a hostage,
5
after which player 1 and player 2 play either a
regular or an altered Trust Game. If player 2 decides not to post the hostage, player 1
and player 2 subsequently play a regular Trust Game. But, if player 2 does decide to
post a hostage, then the payoffs for the case where player 1 subsequently chooses DOWN and
player 2 chooses RIGHT in a Trust Game are changed. In the Hostage Trust Game task
as presented in Groningen the hostage would just be “lost” and player 1 would get his or
her regular S
1
-payoff. In Amsterdam, the hostage would be given to player 1, who would
then end up with S
1
+H
1
instead of just S
1
. Subjects were asked the following questions:
1. About player 1
• If you were player 1, and player 2 had posted the hostage, would you then
choose RIGHT or DOWN?
• If you were player 1, and player 2 had not posted the hostage, would you then
choose RIGHT or DOWN?
2. About player 2
• If you were player 2, would you post the hostage?
5
In Dutch, we used the word “onderpand”, a word that does not have such a strong connotation as
the words “hostage” or “commitment” have in English.
10
Table 2: Payoffs of the 35 (XX IK TEL MAAR 33) different Hostage Trust Games used
in the experiments (H is the value of the hostage, P
1
= P
2
, R
1
= R
2
, number of subject
for each condition in parentheses).
Groningen
Version 1 Version 2 Version 3 Version 4
No. Γ
TG
S
1
-P
i
-R
i
-T
2
H (#) H (#) H (#) H (#)
1. 10-30-60-75 5 (25) – (–) – (–) – (–)
2. 10-30-45-70 – (–) 20 (26) – (–) – (–)
3. 15-40-50-75 – (–) – (–) 5 (27) – (–)
4. 5-35-75-95 – (–) – (–) – (–) 40 (28)
5. 10-30-60-80 5 (25) 25 (26) 15 (27) 10 (28)
6. 15-40-50-85 10 (25) 40 (26) 30 (27) 20 (28)
7. 10-30-45-100 60 (25) 30 (26) 40 (27) 15 (28)
8. 5-35-75-120 25 (25) 10 (26) 35 (27) 50 (28)
Amsterdam
Version 1 Version 2 Version 3 Version 4
No. Γ
TG
S
1
-P
i
-R
i
-T
2
H (#) H (#) H (#) H (#)
3. 15-40-50-75 5 (30) 10 (30) 15 (30) 20 (30)
5. 10-30-60-80 10 (30) 15 (30) 20 (30) 5 (30)
6. 15-40-50-85 10 (30) 15 (30) 20 (30) 5 (30)
7. 10-30-45-100 15 (30) 20 (30) 5 (30) 10 (30)
9. 15-40-50-85 25 (30) 10 (30) 15 (30) 20 (30)
• If you were player 2, would you—given your answer to the previous question—
choose RIGHT or DOWN after player 1 has chosen DOWN?
6
Table 2 summarizes the values of the hostages that were included in the two places
where this task was conducted, Groningen and Amsterdam. As an example, consider
Trust Game 1. This game was used in the Groningen sessions (Version 1) only. The
value of the hostage equaled 5. From the 106 subjects in Groningen, 25 had to decide in
this Hostage Trust Game. Now consider Trust Game number 9. This game was used in
the Amsterdam sessions only (in all versions). The value of the hostage equaled 25, 10,
15, and 30 in Versions 1 through 4. From the 120 subjects in Amsterdam, 30 subjects
participated per version.
The decisions of subjects in the role of player 1 and player 2 constitute the variables
trust (2 posts no hostage) and trust (2 posts hostage) for player 1, and post
hostage and honor trust (given own hostage decision) for player 2. Since subjects
were only presented Hostage Trust Games that were based on Trust Games they had
decided on before, this allows for a direct (within-subjects) comparison between subjects’
behavior in Trust Games and Hostage Trust Games. The value of the hostage constitutes
the variable hostage in the data.
6
Suppose a certain player 2 chose not to post the hostage. Note that this player 2 is then not asked
what he would do after he had posted the hostage, and player 1 had chosen DOWN. The decision not to
ask this to player 2 is defensible, but leads to some difficulties later (see Section 3).
11
In between the choices on Trust Games and those on Hostage Trust Games, subjects
were asked to fill out a questionnaire which included questions about their sex, age,
field of study, knowledge of game theory, whether they were a blood donor, whether
they carried a donor codicil, and composition of their high school final examinations
(whether it included math, physics, economics, French and/or history). The answers
to these questions constitute our control variables male, age, economist, game theory,
blood donor, and codicil. A factor analysis of the composition of subjects’ examination
package constitutes a single variable labeled school. However, the variables blood donor,
game theory, and school did not approach significance in any of the analyses. They were
not incorporated in the final analyses.
One extra complication has to be tackled before reporting the analyses. In the Hostage
Trust Games task, a player 2 was only asked whether they would post the hostage and
given this decision whether they would honor player 1’s trust. If player 2—for a particular
Hostage Trust Game—chose to post the hostage, the data do not contain whether this
player 2 would have honored trust if he had not posted the hostage. That is, subjects did
not answer a question like “You have posted a hostage. But suppose you had not, and
suppose that player 1 trusts you, what would you do?”. The main reason for excluding
this question was that the answers would probably not be very reliable and might confuse
the subject about the consequences of posting a hostage. But we would like an answer to
the question what player 1 would do if he were player 2. It makes sense to just use honor
trust (given own hostage decision) in this case. If player 2 has not posted the
hostage, player 2’s score on honor trust (given own hostage decision) is precisely
what we want to know. If player 2 has posted the hostage and player 2 does not honor
trust (in the Hostage Trust Game), it is likely that player 2 would not have honored trust
if he had not posted the hostage as well. Again, honor trust (given own hostage
decision) seems a good proxy. The only problem arises when player 2 has posted the
hostage and player 2 honors trust. It is then not clear what player 2 would have done,
had he not posted the hostage. Two solutions are possible. Either replace the value of
honor trust (given own hostage decision) by the variable honor trust from the
related Trust Game for those cases where player 2 has not posted the hostage, or just
use honor trust (in the Hostage Trust Game) as it is. Since it turned out to make
no substantial difference in the analyses, all analyses are presented using honor trust
(given own hostage decision).
4 Analyses
All analyses are restricted to hostages that are not bribing (H
1
< P
1
− S
1
) and are not
completely binding (H
2
< T
2
− R
2
). One could argue that all subjects who (when in
the role of player 1) chose to trust player 2 in a Trust Game, can be excluded from the
analysis. Apparently, they saw no problem of trust, so for them there just is no problem of
trust, and it makes no sense to consider cases where a subject was proposed a solution (a
hostage) for a trust problem he does not have. This is a theoretically valid argument, but
empirically it does not make much difference whether they are excluded or not. Adding
a dummy variable that equals 1 if and only if player 1 trusts player 2 in the Trust Game,
does not lead to substantially different results. None of the displayed results are restricted
to those subjects who had a problem of trust. Instead, the complete data set was used.
12
Table 3: Overview of the data.
Number of subjects 226
Number of Trust Games per subject 5
Number of Hostage Trust Games per subject 5
Variable Average St. Dev. Min. Max.
% trust per Trust Game 0.34 0.23 0.1 0.8
% honored trust per Trust Game 0.40 0.21 0.1 0.7
trust 0.26 0.44 0 1
in the Trust Game
honor trust 0.32 0.47 0 1
in the Trust Game
post hostage 0.53 0.50 0 1
in the Hostage Trust Game
trust 0.40 0.49 0 1
(2 posts hostage)
trust 0.18 0.38 0 1
(2 posts no hostage)
honor trust 0.35 0.48 0 1
(given own hostage decision)
age 22.46 2.48 17 55
male 0.61 0.49 0 1
codicil 0.32 0.47 0 1
economist 0.31 0.46 0 1
Groningen 0.47 0.50 0 1
4.1 Convincing player 1
First, we introduce some variable labels. For the Trust Game, we have already defined
risk as
P1−S1
R1−S1
, and temptation as
T2−R2
T2−S1
. This definition is straightforwardly adapted in
the case of Hostage Trust Games. As mentioned before, in a Hostage Trust Game player 1
faces either a Trust Game Γ

TG
or the standard Trust Game Γ
TG
. For Γ

TG
, we redefine
the labels risk (and temptation) as
• risk =
P1−S1−H1
R1−S1−H1
, and
• temptation =
T2−R2−H2
T2−S1−H1−H2
.
In correspondence with the previously given definition, we define labels for the degree of
compensation and the degree of binding.
• compensation =
H1
P1−S1
, and
• binding =
H2
T2−R2
.
13
Table 4: Probit analysis of the probability that player 1 trusts player 2 after hostage
posting. Standard errors are adapted for clustering within subjects.
Independent variable Hypothesis Unstandardized p-value Effect
coefficient
honor trust + 1.51** 0.00 0.53
compensation + 3.17** 0.00 1.06
binding + −0.22 0.12
age ? −0.08** 0.00 −0.03
male ? 0.29* 0.03 0.07
codicil ? 0.33** 0.01 0.13
economist ? −0.42** 0.01 −0.14
Groningen ? 0.76** 0.01 0.49
constant −0.14 0.83
Pseudo-R
2
= 0.27 Number of observations = 943
Log Likelihood = −436.2 Effects at p = 0.28
4.1.1 Compensation and binding - hypothesis 1 through 3
First, consider the effect of the degree of compensation (compensation, or
H1
P1−S1
) versus
the effect of the degree of bonding (binding, or
H2
T2−R2
) on the probability that player 1
trusts player 2 after player 2 has posted the hostage. Table 4 summarizes the results
of a probit analysis of the probability that player 1 trusts player 2 after hostage posting.
Focus on the variables compensation and binding.
Table 4’s main message is clear: compensation has a positive effect on trust, while we
do not find a significant effect for binding. Consequently, the effect of compensation is
larger than the (non-significant) effect of binding. Some objections can be raised against
that conclusion on the basis of this table alone, but this message is essentially correct.
First, it could be argued that since the hostages in Groningen were not compensating at
all (H
1
= 0 for all hostages), comparison of the Groningen data with the Amsterdam data
makes no sense. Table 4 displays a non-significant effect of binding, but a substantial
effect of Groningen, which suggests something fishy may be going on here. We there-
fore estimated the same model on the Amsterdam data only, but found no substantially
different results (analysis not reported). A second argument in favor of careful delib-
eration about the analysis in Table 4, is that it can be argued that the variable honor
trust in the Hostage Trust Game more or less measures the same thing as binding, just
like honor trust in the Trust Game measures more or less the same as temptation,
and should therefore be excluded from the analysis. However, re-estimating the model
without the variable honor trust does not lead to substantially different results either
(analysis not reported). Finally, recall that the idea that the variables compensation and
binding have an effect on the probability that player 1 trusts player 2 at all, is based
on intuitive reasoning only. The more player 1’s potential loss is compensated for if his
trust is abused, the more likely it is that player 1 will trust player 2. Likewise, the more
player 2’s temptation is reduced, the more player 1 will be willing to risk trusting player 2.
14
As a test of these intuitions, the results in Table 4 make sense. On the other hand, what
happens in Trust Games depends on risk and temptation, which was supported in the
analyses. Therefore, it could be argued that the impact of the hostage on trust in Hostage
Trust Games is not brought about by compensation and binding, but by the effect of
the hostage on risk and temptation. This suggests that the previous analyses should
be repeated with risk (
P1−S1−H1
R1−S1−H1
) and temptation (
T2−R2−H2
T2−S1−H1−H2
) included. The effect of
compensation and binding should then disappear, or at least change. However, adding
the variables risk and temptation does not make the effect of compensation completely
disappear, and the effect of compensation remains larger than the (non-significant) effect
of binding. Together, these results imply that Hypotheses 1 and 3 are supported
and that Hypothesis 2 is not supported.
7
Moreover, although one could argue that
the effect of binding or temptation in Hostage Trust Games runs through the variable
honor trust, deleting the variable honor trust from the analysis does not lead to other
insights.
4.1.2 Type revealing hostages - hypotheses 4 and 5
Subsequently, we compare the behavior of player 1 in Trust Games with the behavior
of (the same) player 1 in Hostage Trust Games where no hostage has been posted. If
neglecting to post the hostage does not serve as a signal about player 2’s intent, no
difference between the Trust Games and the Hostage Trust Games should be observed.
The data were manipulated in such a way that each subject represents 10 observations:
5 in Trust Games, and 5 in the related Hostage Trust Games. That is, the variable
trust contains 5 decisions in the role of player 1 in a Trust Game, and 5 decisions in
the role of player 1 in a Hostage Trust Game after the hostage has not been posted. A
dummy-variable hostage game was added to the independent variables to capture the net
added effect of not posting the hostage. Table 5 displays the results of a probit analysis
of the probability that player 1 trusts player 2, comparing Trust Games with Hostage
Trust Games in which the hostage was not posted. Our main interest is in the variables
hostage game and the interaction effect hostage game × Groningen.
Table 5 shows a substantial and significant negative effect of risk on the probability of
player 1 to trust, and no significant effect of temptation. Moreover, there is a small effect
of codicil. The variable Groningen is also significant, just like the variable honor trust.
But, Table 5 also shows that there does exist a negative effect of “not posting a hostage
when you could have”. After player 2 has decided not to post the hostage, player 1 is less
likely to trust player 2 than player 1 would have been in an otherwise similar Trust Game
(at the mean of the independent variables, the estimated decrease in probability is 0.10).
No differences with respect to this effect were found between Groningen and Amsterdam
(the interaction hostage game × Groningen is not significant). Hypothesis 4, that
player 1’s decision whether or not to trust player 2 in the Trust Game Γ
TG
is the same
as player 1’s decision whether or not to trust player 2 in the Hostage Trust Game Γ
HTG
after player 2 has declined to post the hostage, is rejected.
7
The three arguments in this paragraph are countered by three ways of conducting the analysis. (1)
Restricting the analysis to the Amsterdam observations, (2) not including the variable honor trust, and
(3) adding the variables risk and temptation. This results in 8 (2×2×2) possible analyses, all of which
were conducted. The assertion that a hostage’s degree of compensation is more important than its degree
of bonding, remains valid across all these analyses.
15
Table 5: Probit analysis of the probability that player 1 trusts player 2. Comparison be-
tween Hostage Trust Games without a hostage being posted, and Trust Games. Standard
errors are adapted for clustering within subjects.
Independent variable Hypothesis Unstandardized p-value Effect
coefficient
hostage game 0 −0.48** 0.00 −0.10
hostage game 0 0.16 0.24
× Groningen
honor trust + 1.32** 0.00 0.36
risk − −2.08** 0.00 −0.44
temptation − −0.02 0.95
age ? −0.01 0.59
male ? −0.01 0.07
codicil ? 0.30* 0.04 0.07
economist ? −0.21 0.23
Groningen ? 0.87** 0.00 0.20
constant −0.62 0.33
Pseudo-R
2
= 0.29 Number of observations = 1883
Log Likelihood = −668.3 Effects at p = 0.13
For our fifth hypothesis we try to determine in a similar manner if there exists a
positive effect of posting the hostage, other than the change in the payoffs it induces.
We compare—within a Hostage Trust Game—player 1’s decision to trust player 2 in the
subgame that emerges after player 2 has posted the hostage, with player 1’s decision to
trust player 2 in the subgame that emerges after player 2 has not posted the hostage. The
data were transformed so that each subject represents 10 observations instead of 5: per
Hostage Trust Game, the data contain what a subject chose in the role of player 1 after
player 2 has posted the hostage, and what a subject chose in the role of player 1 after
player 2 has not posted the hostage. Each subject made his decisions for 5 Hostage Trust
Games, and each Hostage Trust Game has 2 Trust Games as subgames, which yields the
10 observations per subject. The dummy-variable hostage was posted was added to
capture the effect of having posted the hostage. To really measure the net effect of having
posted the hostage, we need to control for the change in the payoffs that is brought about
by posting the hostage. For that purpose, we added the variables risk and temptation
to the analyses. Table 6 summarizes the results.
As can be seen in Table 6, the effect of the payoffs after controlling for honor trust
is restricted to an effect of risk. A small effect of codicil is once again encountered:
subjects carrying a donor codicil have a higher estimated probability (0.25 compared
to 0.17) to trust player 2. Effects in the same order of magnitude are found for male
and economist. Males are more likely to trust player 2, subjects who study economics
16
Table 6: Probit analysis of the probability that player 1 trusts player 2 in a Hostage
Trust Game. Comparison within Hostage Trust Games. Standard errors are adapted for
clustering within subjects.
Independent variable Hypothesis Unstandardized p-value Effect
coefficient
hostage was posted 0 1.21** 0.00 0.30
hostage was posted 0 −0.97** 0.00 −0.18
× Groningen
honor trust + 1.37** 0.00 0.42
risk − −3.57** 0.00 −0.91
temptation − −0.30 0.31
age ? −0.04 0.08
male ? 0.30* 0.02 0.07
codicil ? 0.29* 0.02 0.08
economist ? −0.35* 0.01 −0.08
Groningen ? 1.35** 0.00 0.37
constant −1.88** 0.00
Pseudo-R
2
= 0.28 Number of observations = 1886
Log Likelihood = −759.5 Effects at p = 0.17
are less likely to trust player 2. The effect of the posting of a hostage once again is
significant and substantial. Distinguishing the effect of the hostage in the Groningen and
Amsterdam experiment, we find an estimated positive effect of the posting of a hostage
on the probability that player 1 trusts player 2 of 0.30 (0.17 to 0.47) in Amsterdam, and
of 0.12 (0.17 to 0.29) in Groningen.
8
The difference between Groningen and Amsterdam
with respect to the effect of the posting of a hostage is somewhat puzzling. It could be
due to the fact hostages, if lost, were paid to player 1. The more noteworthy fact is that
once more a significant effect of the posting of the hostage is encountered. Hypothesis 5
that stated that player 1’s decision whether or not to trust player 2 is solely based on the
payoffs that are brought about by player 2’s decision (and not by the posting of a hostage
itself) is rejected.
Apparently, hostage posting by player 2 triggers trustfulness in player 1. One way
to incorporate that in our measurement models, would be to assume that the posting of
a hostage increases player 1’s subjective probability that player 2 will honor trust, and
this increase is larger for a large degree of compensation of the hostage. In our analyses,
this would imply a hypothesized interaction effect hostage was posted × compensation
on trust. Additional analyses incorporating this interaction effect indeed resulted in a
positive (and significant) interaction effect hostage was posted × compensation. That
8
Incorporating additional interaction effects with Groningen does not reveal other differences between
Groningen and Amsterdam (analyses not reported).
17
Table 7: Probit analysis of the probability that player 1 trusts player 2 if and only if
player 2 posts the hostage. Standard errors are adjusted for clustering within subjects.
Independent variable Hypothesis Unstandardized p-value Effect
coefficient
reduction of risk + 5.00** 0.00 1.26
reduction of temptation + 1.26 0.06
age ? −0.06** 0.01 −0.02
male ? 0.07 0.61
codicil ? 0.19 0.16
economist ? −0.37* 0.03 −0.09
Groningen ? 0.31 0.13
constant −0.36 0.54
Pseudo-R
2
= 0.17 Number of observations = 943
Log Likelihood = −390.4 Effects at p = 0.17
is, the larger the degree of compensation of the hostage, the larger the effect of the mere
posting of a hostage on player 1’s subjective probability that player 2 will honor trust.
4.1.3 Essential hostages - hypothesis 6
A hostage is essential for player 1 if player 1 would not trust player 2 after he neglected
to post a hostage, but would trust player 2 after he did post a hostage. We first created
a dummy-variable essential hostage that equals 1 if player 1 trusts player 2 after
hostage posting and does not trust player 2 after no hostage has been posted. essential
hostage will be the dependent variable in the subsequent analysis. The derivation of
the appropriate measurement model is not without difficulties, but it turns out that
one can use the variables reduction of risk =
P
1
−S
1
R
1
−S
1

P
1
−S
1
−H
1
R
1
−S
1
−H
1
and reduction of
temptation =
T
2
−R
2
T
2
−S
1

T
2
−R
2
−H
2
T
2
−S
1
−H
1
−H
2
as a predictor for whether the hostage is essential
to player 1. See the appendix for the derivation of the measurement model. Table 7
summarizes the results of a probit analysis of essential hostage, including the variables
reduction of risk and reduction of temptation.
As expected, the effect of reduction of risk is significant. The effect of reduction
of temptation is not significant, though removing the variable Groningen from the anal-
yses would lead to a small but significant result of reduction of temptation (analysis
not reported). In any case, the effect of reduction of risk is much larger than the
effect of reduction of temptation. Actually, because of approximations used in the
measurement model, effects of subject characteristics (sex, age, etc.) are ruled out (see
appendix). Nevertheless, there are some effects. Both older subjects and subjects study-
ing econom(etr)ics are less likely to be convinced they should trust player 2 because of
the posting of a hostage. Hypothesis 6 that connects whether a hostage is essential
to the reduction of risk and the reduction of temptation finds partial support. It is
18
Table 8: Probit analysis of the probability that player 2 honors player 1’s trust. Compar-
ison between Hostage Trust Games without a hostage being posted, and Trust Games.
Standard errors are adapted for clustering within subjects.
Independent variable Hypothesis Unstandardized p-value Effect
coefficient
hostage game 0 −0.02 0.98
hostage game 0 −0.00 0.87
× Groningen
temptation − −1.65** 0.00 −0.54
age ? 0.04 0.33
male ? −0.21 0.21
codicil ? −0.12 0.38
economist ? −0.18 0.41
Groningen ? 0.25* 0.21
constant −0.75 0.88
Pseudo-R
2
= 0.09 Number of observations = 1883
Log Likelihood = −1023.2 Effects at p = 0.27
supported with regard to the reduction of risk, but the evidence in favor of the reduction
of temptation is weak.
4.2 Restricting player 2 - hypothesis 7
In Table 5, an effect of the hostage not being posted on the behavior of player 1 was
found. Player 1 is less likely to trust a player 2 who could have posted a hostage but
didn’t. A logical follow-up question is: Should player 1 behave that way? We analyze this
by comparing the behavior of player 2 in Trust Games with the behavior of player 2 in a
Hostage Trust Game in the decision node that emerges after player 2 has decided not to
post the hostage in a Hostage Trust Game, and player 1 has decided to trust. Table 8
shows the results.
To capture the effects of the payoffs, only temptation is included. Its effect on the
probability to honor trust is negative and substantial.
9
Adding other payoff effects, such
as risk, compensation, and binding does not lead to substantially different results. No
effects of age, male, codicil, economist or Groningen are found. Surprisingly, the effect
of not posting a hostage is not significant. Apparently, for the decision whether or not to
honor the trust of player 1 it does not matter for player 2 whether this occurs in a Trust
Game, or in a Hostage Trust Game after they have decided not to post a hostage. This
establishes that it is strange that the mere fact of posting a hostage by player 2 matters
to player 1, since subjects in the role of player 2 do not behave differently in the same
9
The effect of temptation does not differ between Groningen and Amsterdam (analysis not reported).
19
Trust Game conditional on whether or not they posted a hostage before. Adding the
variable post hostage or restricting the analysis to the cases in which player 2 has not
posted the hostage, does also not lead to substantially different results. In all cases, the
non-significance of the effect of hostage game (for Groningen and Amsterdam) remains.
This implies that Hypothesis 7 is supported.
5 Conclusion and discussion
We considered hostage posting as one way to overcome problems of trust. Starting point
were theoretical and experimental results on behavior in Trust Games (Snijders 1996).
The basic results in Trust Games are that trust varies with the risk for player 1 (the
trustor) and the temptation for player 2 (the trustee). A hostage that can overcome
a problem of trust should therefore either decrease the risk or reduce the temptation.
Stated otherwise, a hostage that can overcome a problem of trust should either promise
to compensate if trust is abused, or sufficiently reduce the probability that trust will be
abused. Trust Game analyses also showed that (for the trustor) risk is more important
than temptation. Hence, one could expect that a hostage that decreases the risk is
more likely to overcome a problem of trust than a hostage that reduces the temptation.
This is indeed what we find; our analyses confirm that the degree of compensation is
more important than the degree of reduction of incentive to abuse trust. From a game-
theoretic point of view this may appear somewhat weird. The trust problem is caused by
the fact that player 2 has an incentive to abuse. One might think that it is therefore of
major importance to decrease this incentive for player 2 to abuse. But apparently, this
is not how it works. Compensating player 1 to some degree works much better to get
player 1 to trust. The kind of behavior observed in our experiments shows subjects to be
somewhat myopic, in the sense that they do not appreciate solutions to a trust problem
that work on the other player’s incentives. On the other hand, one could argue that this
kind of behavior makes perfect sense. For a player 1 it is much easier to assess the use of
some compensation for him, than it is to assess the use to him of decreasing incentives of
player 2.
Hostage posting that credibly reveals that the person who makes this commitment is
the kind of person who would never abuse trust, was labeled a type revealing hostage.
That is, it is neither the compensation in case trust gets abused, nor the bonding of
the trustee to honor trust that is brought about by posting the hostage, that makes the
potential trustor trust. It is the mere fact that this person has made this commitment,
that makes the potential trustor trust. We argued that hostages cannot be type revealing
in Hostage Trust Games. The main factor that carries this argument is that posting a
hostage is always profitable, whatever kind of player 2 you are. Empirically, this was
investigated by comparing the subgames of the Hostage Trust Games that emerge after
no hostage has been posted by player 2. If “player 2 posts no hostage” has no intrinsic
value, what remains is a standard Trust Game. And indeed we find that the mere fact
of posting a hostage does not go together with a player 2 being more likely to honor
trust. What determines whether player 2 honors trust is again temptation (
T2−R2
T2−S1
), and
nothing else. In contrast with this, the analyses suggest that player 1 sees some kind of
negative intrinsic value in “not posting a hostage when you could have” (a negative effect
on the probability to trust of 0.11). Comparing subgames within Hostage Trust Games
20
confirmed this result (a positive effect of posting the hostage of 0.10 in Groningen, and
0.27 in Amsterdam). Though the effects are not that large, the theory is evidently refuted
here: posting a hostage works to build trustfulness, even though a theoretical argument
is hard to imagine.
10
And although player 1 considers posting a hostage as a signal for
trustworthiness, the fact that posting a hosting does not change player 2’s behavior shows
that the hostage posting signal is deceptive.
Finally, a word of advice for those who want others to trust them. First, post the
hostage when you can, even if it hardly affects the payoffs. “Not doing something to
overcome the problem” reflects badly on your trustworthiness. Second, if you have a
choice, post a hostage that provides some compensation to the trustor. Whether it binds
yourself to honor trust is of lesser importance to the trustor. Similar advice can be given
to those who have to decide whether or not to trust someone. First, do not mind too
much about whether or not the other person posted the hostage or not, but keep your eye
on what the hostage—if offered—does with the incentive for the trustee. In any case, do
not be taken in by the mere fact that a hostage is being posted. This “signal of benign
intent” does not induce a moral imperative to honor trust.
References
Beard, T. R. and R. O. Bell Jr (1994). Do People Rely on the Self-Interested
Maximization of Others. Management Science 40, 252–262.
Becker, G. S. (1991). A Treatise on the Family. Harvard: Harvard University Press.
Enlarged edition.
Buskens, V. (1999). Social Networks and Trust. Amsterdam: Thela Thesis.
Cho, I. and D. Kreps (1987). Signalling Games and Stable Equilibria. Quarterly Jour-
nal of Economics 102, 179–221.
Crawford, R. and J. Sobel (1982). Strategic Information and Transmission. Econo-
metrica 50, 1431–1452.
Dasgupta, P. (1988). Trust as a Commodity. In D. Gambetta (Ed.), Trust: Making
and Breaking Cooperative Relations, pp. 49–72. Oxford: Blackwell.
Davis, D. D. and C. A. Holt (1993). Experimental Economics. Princeton, NJ: Prince-
ton University Press.
Farrell, J. (1987). Cheap Talk, Coordination, and Entry. Rand Journal of Eco-
nomics 18, 34–39.
Frank, R. H., T. Gilovich, and D. T. Regan (1993). Does Studying Economics
Inhibit Cooperation? Journal of Economic Perspectives 7, 159–171.
Gouldner, A. W. (1960). The Norm of Reciprocity: A Preliminary Statement. Amer-
ican Sociological Review 25, 161–178.
Kahneman, D. and A. Tversky (1979). Prospect Theory: An Analysis of Decision
under Risk. Econometrica 47, 263–291.
10
Another indication that hostage posting affects the trust decision is the fact that the goodness of fit
of the theoretical model that predicts trust after hostage posting on the basis of risk and temptation
is worse for Hostage Trust Games than for Trust Games.
21
Keren, G. and W. Raub (1993). Resolving Social Conflicts through Hostage Post-
ing. Theoretical and Empirical Considerations. Journal of Experimental Psychology,
General 122, 429–448.
Kreps, D. M. (1990). Corporate Culture and Economic Theory. In J. Alt and
K. Shepsle (Eds.), Perspectives on Positive Political Economy, pp. 90–143. Cam-
bridge: Cambridge University Press.
McLennan, A. (1985). Jusitfiable Belief in Sequential Equilibrium. Econometrica 50,
277–323.
Rasmusen, E. (1994). Games and Information: An Introduction to Game Theory (2nd
ed.). Oxford: Blackwell.
Raub, W. and J. Weesie (1990). Reputation and Efficiency in Social Interactions: An
Example of Network Effects. American Journal of Sociology 96, 626–654.
Rosenthal, R. (1981). Games of Perfect Information, Predation Pricing, and the Chain
Store Paradox. Journal of Economic Theory 24, 92–100.
Schelling, T. C. (1960). The Strategy of Conflict. Harvard: Harvard University Press.
Schelling, T. C. (1984). Choice and Consequence. Harvard: Harvard University Press.
Snijders, C. (1996). Trust and Commitments. Amsterdam: Thesis Publishers.
Snijders, C. and G. Keren (1999). Determinants of trust. In D. V. Budescu,
I. Erev, and R. Zwick (Eds.), Games and Human Behavior, pp. 355–385. Mahwah,
NJ: Lawrence Erlbaum.
Williamson, O. E. (1985). The Economic Institutions of Capitalism. New York: The
Free Press.
Appendix: Measurement model for the analysis of hy-
pothesis 6
We refer to the guilt model to derive our measurement model for Hypothesis 6. According
to the guilt model, player 1 assesses the hostage to be essential if and only if the subjectively
expected probability that player 2 honors trust (p
2,1
) is not large enough to make player 1
trust in the Trust Game, but large enough to make player 1 trust in the Hostage Trust Game:
p
2,1
(¬H) <
P
1
−S
1
R
1
−S
1
and p
2,1
(H) >
P
1
−S
1
−H
1

2
(T
2
−S
1
−H
1
−H
2
)
R
1
−S
1
−H
1

2
(T
2
−S
1
−H
1
−H
2
)
. Assume first that
p
ijg
2,1
= ζ
i

¸
T
2
−R
2
−H
2
T
2
−S
1
−H
1
−H
2

g
+
ijg
1
,
with ζ
i
a (vector of) characteristics of player i (like sex or age) and
ijg
1
∼ N(0, σ
2
) the
independently distributed error terms. It follows that
Pr {Hostage is essential} =
= Pr

p
2,1
(¬H) <
P
1
−S
1
R
1
−S
1
and p
2,1
(H) >
P
1
−S
1
−H
1

2
(T
2
−S
1
−H
1
−H
2
)
R
1
−S
1
−H
1

2
(T
2
−S
1
−H
1
−H
2
)

= Pr

ijg
1
(¬H) <
P
1
−S
1
R
1
−S
1
−ζ
i
−β
¸
T
2
−R
2
T
2
−S
1

g
and
22

ijg
1
(H) >
P
1
−S
1
−H
1

2
(T
2
−S
1
−H
1
−H
2
)
R
1
−S
1
−H
1

2
(T
2
−S
1
−H
1
−H
2
)
−ζ
i
−β
¸
T
2
−R
2
−H
2
T
2
−S
1
−H
1
−H
2

g
¸
.
To facilitate statistical estimation, we assume that
ijg
1
(¬H) =
ijg
1
(H) =
ijg
1
. That is, the
error that is made in predicting player 1’s move in the subgame that arises after player 2 has
decided not to post the hostage equals the error that is made in predicting player 1’s move in
the subgame that arises after player 2 has posted the hostage. In other words, we assume that
the error that is made in predicting player 1’s behavior is solely due to player 1’s “mood” (and
not to, say, the specific Trust Game). Given that assumption,
Pr {Hostage is essential} =
Pr

P
1
−S
1
−H
1

2
(T
2
−S
1
−H
1
−H
2
)
R
1
−S
1
−H
1

2
(T
2
−S
1
−H
1
−H
2
)
−ζ
i
−β
¸
T
2
−R
2
−H
2
T
2
−S
1
−H
1
−H
2

g
<
ijg
1
<
P
1
−S
1
R
1
−S
1
−ζ
i
−β
¸
T
2
−R
2
T
2
−S
1

g
¸

Φ

1
σ
P
1
−S
1
R
1
−S
1

ζ
i
σ

β
σ
T
2
−R
2
T
2
−S
1


Φ

1
σ
P
1
−S
1
−H
1

2
(T
2
−S
1
−H
1
−H
2
)
R
1
−S
1
−H
1

2
(T
2
−S
1
−H
1
−H
2
)

ζ
i
σ

β
σ
T
2
−R
2
−H
2
T
2
−S
1
−H
1
−H
2

,
with Φ the standard normal cumulative distribution function.
11
Even given the simplifying
assumption on the error terms, the extra Φ-term that pops up here causes the statistical analysis
to be non-trivial. Although the probability that the hostage is essential increases with
Φ

1
σ
P
1
−S
1
R
1
−S
1

ζ
i
σ

β
σ
T
2
−R
2
T
2
−S
1

−Φ

1
σ
P
1
−S
1
−H
1

2
(T
2
−S
1
−H
1
−H
2
)
R
1
−S
1
−H
1

2
(T
2
−S
1
−H
1
−H
2
)

ζ
i
σ

β
σ
T
2
−R
2
T
2
−S
1

,
this need not imply that it increases with, say,
P
1
−S
1
R
1
−S
1
. We simplify further by assuming
that the Φ(.)-terms are in the range [0.08, 0.4] (a more detailed analysis of this issue and an
explanation why this is reasonable can be found in Snijders 1996: 162–164. On that range, a
linear approximation of Φ performs reasonable. For instance, Φ(x) ≈ 0.28x + 0.44 (the first
order Taylor expansion in x = Φ
−1
(0.2)). Since this implies that Φ(x) −Φ(y) ≈ Φ(x−y) −0.44,
it follows that
Pr {Hostage is essential} =
= Φ

−3.96 + 1.99
T
2
−R
2
T
2
−S
1
+ 3.9
P
1
−S
1
R
1
−S
1


Φ

−3.96 + 1.99
T
2
−R
2
−H
2
T
2
−S
1
−H
1
−H
2
+ 3.9
P
1
−S
1
−H
1
R
1
−S
1
−H
1


= Φ

1.99
T
2
−R
2
T
2
−S
1
+ 3.9
P
1
−S
1
R
1
−S
1
−1.99
T
2
−R
2
−H
2
T
2
−S
1
−H
1
−H
2
−3.9
P
1
−S
1
−H
1
R
1
−S
1
−H
1

−0.44.
11
The “≈” is necessary since it need not be the case that
P
1
− S
1
− H
1

2
(T
2
− S
1
−H
1
−H
2
)
R
1
− S
1
− H
1
+ γ
2
(T
2
− S
1
− H
1
−H
2
)
− ζ
i
−β

T
2
− R
2
− H
2
T
2
−S
1
−H
1
− H
2

g
<
P
1
− S
1
R
1
− S
1
− ζ
i
−β

T
2
−R
2
T
2
− S
1

g
.
23
= Φ(1.99 reduction of risk + 3.9 reduction of temptation) −0.44.
A winding detour, but a relatively straightforward result: the larger the reduction of risk the
hostage induces, the more likely that the hostage is considered essential. And likewise, the larger
the reduction of temptation a hostage induces, the more likely that the hostage is considered
essential. The former effect is larger than the latter. A more pragmatic way of dealing with the
statistics would be to conclude—on the basis of the above analysis—that reduction of risk
and reduction of temptation must have something to do with essential hostage.
24

Master your semester with Scribd & The New York Times

Special offer for students: Only $4.99/month.

Master your semester with Scribd & The New York Times

Cancel anytime.